58 Late-Breaking: Vision-based ResNet-50-CNN and YOLOv8 models for predicting behavioral patterns under heat stress in lactating Holstein and Jersey Cows.

Abstract Lactating dairy cows, including Holstein and Jersey breeds, are particularly vulnerable to heat stress, which elicits distinct behavioral adaptations such as increased respiration, standing duration, decreased feed intake, elevated water consumption, and heightened restlessness. These behavioral responses can significantly compromise animal health, reduce productive efficiency, and impair overall welfare. While these behavioral adaptations serve as critical indicators of heat stress, their detection in real-world cattle herd settings remains challenging and prone to observer bias. Therefore, integration of advanced digital and time-efficient technologies in livestock systems is required to enhance monitoring, reduce handling, and improve welfare perceptions. Machine learning driven computer vision systems are emerging for behavioral monitoring in cattle, yet predictive applications remain limited for lactating cows exhibiting heat stress responses. This study aimed to develop artificial-intelligence-based models using Residual Network-50 Convolutional Neural Networks (ResNet-50-CNN) and You Only Look Once (YOLOv8) to predict differential behavior patterns, during heat stress conditions, among Holstein and Jersey breeds of lactating dairy cows. Twenty-four lactating Holsteins and twelve Jerseys were housed in free stalls during mid-summer (Starkville, Mississippi State, USA) and were provided with ad libitum access to feed and water during a 14-day period. Behavioral observations, i.e., standing, lying, and eating, among Holstein and Jersey lactating cows were captured via twelve stationary ReoLink 5MP PoE IP cameras during four daily intervals (06:00-08:00, 13:00-15:00, 18:00-20:00, and 01:00-03:00), from day 0 to 14, respectively. Raw video footage was processed, and unique frames were extracted at 30-second intervals for each behavior type and annotated for individual animal behaviors i.e. “Standing”, “Lying”, and “Eating”, in each frame using Roboflow software. A total of 1,120 extracted frames (730 for Holstein and 390 for Jersey) were used for training and validation (20%, 100 epochs) using the raw videos from the adaptation days. Additionally, live videos of the four-heat stress (THI>72) trial days were reserved for independent model performance testing. Our trained model, upon validation, showed high accuracy of 0.979 and precision of 0.965 in classifying multi-class behavior in a single frame. When evaluated on unlabeled independent raw videos, it maintained robust performance with an accuracy of 0.892, precision of 0.895, and recall rate of 0.874 in predicting behavior phenotypes and correctly differentiating breed types. Our findings demonstrate that the CNN and YOLOv8 models can effectively distinguish behavior classes and duration across multiple cows and breeds, with promising potential for future applications in detecting heat stress, health indicators, and targeted behavioral traits.